selected work / anonymized
Work across AI platforms, agents, automation, and security
Most of my strongest work sits behind NDAs or inside regulated environments. This page describes the types of systems, decisions, and outcomes without exposing client data.
If you need exact references, I can discuss relevant context privately where disclosure is appropriate.
Banking AI platform and production LLM workflows
A regulated banking environment needed a path from AI experiments to maintainable production systems.
What I did
- Owned AI direction across platform, ML/AI product lifecycle, and production delivery.
- Shaped architecture for LLM systems, evals, guardrails, observability, cost control, and CI/CD.
- Coordinated product, analytics, ML, data engineering, and domain stakeholders.
Outcome
A stronger production operating model for AI work: clearer ownership, platform thinking, review discipline, and reusable delivery patterns.
Enterprise AIBankingLLM PlatformEvalsGovernance
Agentic diagnostics and career-orientation system
An education workflow needed an AI system that could support psychodiagnostics and professional orientation.
What I did
- Designed an agentic system around structured diagnostics, context, recommendations, and workflow state.
- Connected product requirements with LLM behavior, data model, and operational handoff.
- Worked through reliability and user-facing constraints instead of treating it as a generic chatbot.
Outcome
A focused agentic workflow with clearer task boundaries, reusable decision logic, and a path toward practical deployment.
AgentsEdTechDiagnosticsWorkflow Design
Corporate AI workshops and enablement
Enterprise teams needed to understand what LLMs and agents can do, where they fail, and how to choose realistic first projects.
What I did
- Prepared and delivered AI, LLM, and agent workshops for corporate teams.
- Translated model capabilities into operational patterns: context, tools, evals, security, and ownership.
- Helped teams separate valuable use cases from demo-driven noise.
Outcome
Teams left with more realistic AI roadmaps, fewer vague pilots, and a stronger vocabulary for production decisions.
WorkshopsAI ReadinessAgentsEnterprise Enablement
AI automation, bots, Mini Apps, and internal tools
Small teams and operators needed pragmatic automation without waiting for a large internal platform.
What I did
- Built Telegram bots, Mini Apps, n8n pipelines, and internal workflows around real business processes.
- Integrated external APIs, structured forms, notifications, and human review points.
- Balanced speed with maintainability, observability, and data boundaries.
Outcome
Faster operational workflows and clearer path from manual process to AI-assisted execution.
AutomationTelegramn8nInternal ToolsAI Workflows
LLM security and prompt-injection research
Tool-using agents and RAG systems introduce risk beyond ordinary chatbot behavior.
What I did
- Published on LLM-injection attack types and protection patterns.
- Analyzed prompt injection, tool misuse, data leakage, context contamination, and MCP/tooling risk.
- Translated security concerns into review checklists, evals, and architecture constraints.
Outcome
A sharper security angle for production AI work: not only model behavior, but permissions, tools, context, traces, and human approvals.
AI SecurityPrompt InjectionMCPTool RiskGuardrails